TCS1: Trajectory Candidate Selection
Status: closed negative on development; holdout sealed.
Hypothesis
The qualified 9B depth-one owner, blind depth-two reuse, and independently
trained direct-rewrite owner solve different development identities. Their
fixed union is 615/1,289: math 228, logic/science 368, and code 19.
TCS1 asks whether model-owned candidate assessment can recover a material part
of that complementarity without a verifier, answer key, task label, external
model, or fieldwise trajectory mixing at inference.
Frozen sequence
- Build train-only three-candidate groups from the original base and expert attempts plus the actual IDR1 predecessor for 5,824 source identities.
- Build the source-disjoint 1,289-row development pool from complete depth-one, depth-two, and direct-rewrite trajectories. Holdout remains unopened.
- Fit a CPU shape-only control that cannot read lineage, task labels, gold, correctness, or peer outcomes. It may read only candidate text, question text, generated length, and exhaustion metadata.
- Only after the CPU receipt is immutable may one semantic hidden-state scorer be frozen. It must use the same candidate pool and whole-trajectory argmax.
Semantic promotion gate
The future semantic scorer must satisfy every condition:
- at least
603/1,289, fourteen above depth one and at least 53% of the 26-answer oracle opportunity; - at least ten answers above the frozen shape-only control;
- math at least 223, logic/science at least 349, and code at least 17;
- at least eight-answer loss when candidate contents are deterministically permuted while questions and candidate positions remain fixed;
- no task/benchmark label, lineage identifier, correctness field, verifier, answer key, or external model at inference; and
- complete candidate, parameter, FLOP, latency, generated-token, truncation, and protected-checkpoint receipts.
A pass opens one source-disjoint holdout selection. A failure closes this exact candidate-commit route without prompt, width, seed, threshold, pool, or duration rescue. The claim is practical coherent trajectory commitment, not a new reasoning primitive.
CPU control result and semantic owner
The immutable candidate set contains 17,472 train candidates over 5,824
identities and 3,867 development candidates over 1,289 identities. Train and
development SHA-256 values are 622aa19e...9022 and a216928d...5048.
Shape-only job 747538 scores 579/1,289: it repairs one depth-one error and
breaks eleven successes. Domains are math 214, logic/science 348, and code
17. It therefore fails every promotion condition and rules out superficial
completion-shape selection as the source of the oracle gap.
Exactly one semantic scorer is now frozen. The pinned Qwen3.5-9B IDR1 adapter
reads the raw source question and each complete candidate independently using
the fixed correctness-assessment instruction. A 512-wide scalar head and the
existing LoRA tensors train for 256 updates, eight groups per update, LR
2e-6/2e-4, seed 2026080903, and maximum context 2,048. Training uses only
the 5,824 train identities and balances task/outcome-pattern strata. Loss is
balanced binary correctness plus within-group positive-over-negative ranking.
Inference takes an argmax over three complete trajectories; scores or fields
are never averaged. The source checkpoint remains immutable.
One one-update mechanics job may run first. If finite, non-OOM, hash-safe, and fully receipted, it unlocks the single 256-update fit above. There is no shuffled-label, width, prompt, layer, seed, or duration family. Development is read once after training; holdout remains sealed behind the frozen gate.
Development result
Mechanics job 747585 passed. The sole semantic fit/evaluation job 747589
then completed all 256 updates and 2,048 group presentations in 960.31 seconds
on one H100. It charged 2,214,631 training tokens, used 4,942,337 trainable
parameters (2,704,896 existing adapter tensors plus a 2,237,441-parameter
head), peaked at 33,433,896,960 GPU bytes, and preserved the protected IDR1
adapter bit-identically. One training presentation truncated; no development
candidate truncated.
The semantic owner selects 565/1,289, below both depth one (589) and the
shape-only control (579). It repairs 14 depth-one errors but breaks 38 prior
successes. Domains are math 205, logic/science 342, and code 18; only the
code floor passes. Deterministically permuting candidate contents reduces the
score by only six answers (565 -> 559), below the frozen eight-answer causal
margin. The selector chooses depth one 1,080 times, depth two 127 times, and
direct rewrite 82 times.
Every conjunctive promotion gate except code fails. Report SHA-256 is
324d59228b96aa9ffc2cbed94467a4156ec46455d0c07a7b907e9664c5ea88ac;
selector checkpoint SHA-256 is
f9055f87c8e9a9c41b3bf8a40e71acccfdb60c1ec6c9a8acc29893f14d9d375d.
The exact result receipt is
docs/research/SHOHIN_TCS1_DEVELOPMENT_RESULT.json.
Decision: close exact TCS1 without prompt, width, seed, threshold, candidate pool, or duration rescue. Holdout remains sealed. The three-way oracle proves that complementary trajectories exist, but neither superficial shape nor the trained hidden-state scorer can identify them without destroying too many correct first revisions. The next useful intervention must improve proposal or revision generation, or train a genuinely joint correction process; another post-hoc candidate selector is not authorized.